Skip to main content

Module interval

Module interval 

Source
Expand description

The readings’ 90% intervals (ADR-257 §1, ADR-308).

A reading’s bound is the most its error can be, held on the tuning set; a 90% interval is narrower, and says how far off the reading usually is. Each comes from the reading’s own bound: the interval runs from the reading less Shares::high of its bound to the reading less Shares::low of it. The two shares are the 5% and 95% points of the reading’s error as a share of its bound, (reading − truth)/bound, measured on the synthetic flights’ own calibration set (2,000 flights on a seed of their own, each flight’s own logger’s log) and written here to three significant figures. The point from each end is the ⌊0.05 (n + 1)⌋-th of the n sorted shares, the split-conformal rank (V. Vovk, A. Gammerman and G. Shafer, Algorithmic Learning in a Random World, Springer, 2005, ch. 2; J. Lei et al., “Distribution-free predictive inference for regression”, JASA 113(523), 1094–1111, 2018). cargo xtask synthetic fits them again and fails if they differ from the values here, then holds the intervals to the coverage test of ADR-257, calibrated intervals, on the two test sets.

How far to trust it. Over flights drawn as the synthetic ones are, a reading’s interval holds the truth about 90% of the time and misses about 5% on each side. The 90% is over their days, spread evenly within 15 K of the standard atmosphere, and the misses aren’t: most of the height’s error is the day’s air, so no height missed on a day within 10 K of standard, while a third of the days warmer than that read low enough to leave the truth above the interval, the flattering side for a waiver (#1055). The top speed and the mean descent rate, taken from the same height, miss the same way, the flattering side for a fin’s margin and for how gentle the descent was (#1061); for flights at or past the top speed’s transonic mark, only its low readings are given. The times lean by logger: on the AltOS logs, the apogee time’s interval misses late about one time in six or seven (#1056), and liftoff, landing and the flight time lean on others (#1062). The top acceleration’s interval leans by logger too: on the AltOS logs the truth misses about one time in five, and on the 500 Hz IMU’s it never did (#1070). The canopies’ rates miss as the height does, by the day’s air. The top speed is tested only where it is given, below super::TRANSONIC_SPEED_M_S. A real flight can lie outside what the sets draw, as a day more than 15 K from standard does; its interval then says less than 90%. No real flight has yet checked one (evidence row G34).

A floor (ADR-308) states only a lower end, the reading less its 95% share of its bound (FloorShares), as its truth may lie any amount above it. A burnout from the top speed’s peak fits its share on its own floors. How far past a clip a top acceleration goes is the design’s, and no log says: a first fit on tops as flown put the lower end above the reading, and one design whose tops only just passed the clip left 21 of 22 truths below it. So its share is fitted on every top the reader can call clipped and the part may not have clipped: from where a gain s high and the part’s bias carry a top up to the reader’s threshold, a quarter g under the range, to 1/(1 − s), past which every part clips. A truth lies below the lower end only on a top under the clip that the gain carried up to it; a top truly past the clip lies above the reading. An apogee floor states none, nor a top whose burn the log ends in: no synthetic set reads one (#1071).

Structs§

FloorShares
A floor’s error, (reading − truth)/bound, on the calibration set. A floor’s truth may lie any amount above it, so only its lower side is stated.
Interval
A 90% interval on a reading: on the synthetic flights, the truth lay below Interval::low about 5% of the time and above Interval::high about 5%.
Shares
A reading’s error, (reading − truth)/bound, at its 5% and 95% points on the calibration set.

Constants§

APOGEE_HEIGHT_SHARES
The apogee height’s shares of its bound (Apogee::bound_m), fitted on the calibration set.
APOGEE_TIME_SHARES
The apogee time’s shares of its bound (Apogee::time_bound_s), fitted on the calibration set.
BURNOUT_FLOOR_SHARES
A burnout floor’s shares of its bound (Burnout::bound_s), fitted on the calibration set.
BURNOUT_SHARES
Burnout’s shares of its bound (Burnout::bound_s), where it isn’t a floor, fitted on the calibration set.
DESCENT_RATE_SHARES
The mean descent rate’s shares of its bound (Landing::mean_descent_rate_bound_m_s), fitted on the calibration set.
DROGUE_RATE_SHARES
The drogue’s rate’s shares of its bound (Descent::bound_m_s), fitted on the calibration set.
FLIGHT_TIME_SHARES
The flight time’s shares of its bound (FlightTime::bound_s), fitted on the calibration set.
LANDING_SHARES
Landing’s shares of its bound (Landing::bound_s), fitted on the calibration set.
LIFTOFF_SHARES
Liftoff’s shares of its bound (Liftoff::bound_s), fitted on the calibration set.
MAIN_OPENING_SHARES
The main’s opening’s shares of its bound (Opening::bound_s), fitted on the calibration set.
MAIN_RATE_SHARES
The main’s rate’s shares of its bound (Descent::bound_m_s), fitted on the calibration set.
TOP_ACCELERATION_FLOOR_SHARES
A top acceleration floor’s shares of its bound (MaxAcceleration::bound_m_s2), fitted on the calibration set’s flights with their accelerometers clipping at 8 g and their tops scaled to anywhere the reader can call them clipped and the part may not have clipped: how far past a clip a top goes depends on the design, and no log says.
TOP_ACCELERATION_SHARES
The top acceleration’s shares of its bound (MaxAcceleration::bound_m_s2), where it isn’t a floor, fitted on the calibration set.
TOP_SPEED_SHARES
The top speed’s shares of its bound (MaxSpeed::bound_m_s), fitted on the calibration set.
Z_90
The standard normal’s 95% point: a 90% interval’s half-width over its standard error.

Functions§

fit
The 5% and 95% points of shares, each the tail_rank-th from its end. None when there are fewer than 20 shares or any isn’t finite.
fit_floor
A floor’s 95% point, the tail_rank-th from the top, and its median. None when there are fewer than 20 shares or any isn’t finite.
tail_rank
The rank from each end of n sorted shares that is the 5% or 95% point, ⌊0.05 (n + 1)⌋, counted from 1; at least 1.